CN117709536B - A method and system for accurate prediction of industrial processes using deep recursive random configuration networks - Google Patents

Abstract

The present invention belongs to the field of dynamic modeling technology, and in particular, relates to a method and system for accurately predicting industrial processes using a deep recursive random configuration network. The method can use stored historical information for modeling, and only inputs data at the current moment, without the need to obtain all order delay data. Under the condition that the root mean square error of the output of the neural network model is reduced, suitable neurons are found as candidate neurons in the reserve pool through the inequality constraints of the random configuration algorithm. The connection weights of the newly added neurons are redistributed, and only the network connection weights of the newly added neurons to themselves and the original neurons are retained, and the network connection weights of the original neurons to the newly added neurons are set to zero, that is, there is no connection. After determining the newly added neurons, the output weights are calculated using the least squares method to determine whether the reserve pool of the current layer has been completed. Then configure the reserve pool of the next layer, and determine whether the network has been completed by the maximum number of layers, the maximum allowable number of neurons in the reserve pool, and the maximum allowable output error.

Description

A method and system for accurate prediction of industrial processes using deep recursive random configuration networks

Technical Field

The present invention belongs to the technical field of dynamic modeling, and in particular relates to a method and system for accurately predicting industrial processes using a deep recursive random configuration network.

Background technique

Complex systems in industrial processes have multivariable dynamic evolution behaviors. Each input variable has a certain delay, that is, the effect of the current input on the system will be reflected after a period of time. However, due to the influence of process environment, external disturbances and other factors, it is difficult for us to capture all order delay information involving input variables, and the input order at different times is also constantly changing. How to effectively use known variable information to accurately model such nonlinear dynamic systems with uncertain orders is very important.

Through the above analysis, the problems and defects of the prior art are as follows:

(1) Recurrent neural networks and long short-term memory networks can store some historical information because of the feedback connections between their neurons, and can be used to solve the problem of modeling nonlinear dynamic systems with uncertain orders. However, these two networks need to adjust the network connection weights and biases according to the error gradient descent method during training. Therefore, the training process is computationally intensive and time-consuming, and it is easy to fall into local optimality, resulting in problems such as gradient disappearance or explosion, and the implementation efficiency is low.

(2) There are also feedback connections between neurons in the echo state network. Compared with the recurrent neural network, the echo state network calculates the output weights through the least squares method, and the input weights and the reserve pool connection weights are randomly given and no longer change. However, the echo state network has the disadvantages of blind selection of network structure, sensitive weight parameters, and inability to guarantee the global approximation properties of the model. In addition, it is unable to adaptively adjust the range of weights and biases according to the input data, which affects the prediction accuracy of the model.

(3) The reserve pool of the recursive random configuration network also uses random connections, introduces inequality constraints, and adds reserve pool neurons in a supervised manner, which can effectively avoid the above problems. However, compared with the multi-layer network model, when the total number of neurons is similar, the computational complexity of the single-layer echo state network and the recursive random configuration network is larger, and the model training speed is slow.

In the industrial field, the problems and defects of the above technologies lead to the following problems:

1. Lack of real-time performance: Since recursive neural networks and long short-term memory networks (LSTM) need to adjust network connection weights and biases according to the error gradient descent method during training, the training process is computationally intensive and time-consuming, and is not suitable for industrial applications that require fast response, such as real-time monitoring and fault detection.

2. Low prediction accuracy: The echo state network and recursive random configuration network have low prediction accuracy due to problems such as sensitive weight parameters and the inability to guarantee the global approximation properties of the model. They cannot meet the requirements of industrial applications that require high-precision predictions, such as quality control and energy system load forecasting.

3. Low resource utilization efficiency: Compared with multi-layer neural networks, single-layer recursive random configuration networks have higher computational complexity and slower model training speed when the total number of neurons is similar, resulting in low resource utilization efficiency.

4. Insufficient stability and robustness: Since recursive neural networks and long short-term memory networks are prone to falling into local optimality, resulting in problems such as gradient disappearance or explosion, the model's stability and robustness are insufficient, which poses risks for industrial systems that need to operate stably under various working conditions.

5. Difficulty in adjusting model parameters: Since these models cannot adaptively adjust the range of weights and biases according to input data, it is difficult to adjust the models and they cannot quickly adapt to changes in industrial systems, affecting the operating efficiency and effectiveness of the system.

Summary of the invention

In view of the problems existing in the prior art, the present invention provides a method and system for accurately predicting industrial processes using a deep recursive random configuration network, the purpose of which is to solve the problem of modeling nonlinear dynamic systems with uncertain orders, and to improve the overall training speed and prediction accuracy of the model. The deep recursive random configuration network can use the stored historical information for modeling, and only input the data at the current moment, without obtaining all the order delay data. Under the condition that the root mean square error of the output of the neural network model is reduced, the inequality constraints of the random configuration algorithm are used to find suitable neurons as candidate neurons in the reserve pool. Different from the original random connection of the reserve pool, the deep recursive random configuration network redistributes the connection weights of the newly added neurons, and only retains the network connection weights of the newly added neurons to themselves and the original neurons, and sets the network connection weights of the original neurons to the newly added neurons to zero, that is, no connection. After determining the newly added neurons, the output weights are calculated using the least squares method, and the maximum allowable number of neurons in the reserve pool and the maximum allowable output error are used to determine whether the reserve pool of the current layer is completed. Then, the same steps are followed to configure the next layer of the reserve pool, and the process continues to move forward to deeper reserve pools. The maximum number of layers, the maximum allowable number of neurons in the reserve pool, and the maximum allowable output error are used to determine whether the network has been constructed.

The present invention is implemented as follows: a method for accurately predicting industrial processes using a deep recursive random configuration network. This technical solution adopts an accurate prediction method based on a deep recursive random configuration network (DeepRSCN). Through data preprocessing, network layer initialization, random configuration weights and biases, selection and optimization of reserve pool neurons in combination with inequality constraints, and a projection algorithm using real-time data updates, dynamic system modeling with high adaptability and high prediction accuracy is achieved. This solution reduces the reliance on a large amount of label data, significantly improves computational efficiency, has good scalability and real-time performance, and is suitable for dynamic modeling of complex systems that require fast response and high-precision prediction.

Further, including:

S1, data preprocessing;

S2, deep recursive random configuration network initialization;

S3, randomly configures weights and biases, and uses a new method to construct candidate neurons in the reserve pool;

S4, supervised selection of candidate neurons based on inequality constraints;

S5, determine the best candidate neurons and add them to the reserve pool;

S6, calculate the output weight of the current layer network by combining the target output and the reserve pool output;

S7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3 to S6 until the error requirement is met or the maximum allowable number of neurons in the reserve pool of the current layer is reached;

S8, when the error requirement is not met, enter the next layer and configure the reserve pool neurons of each layer in sequence according to S3 to S7 until the error requirement is met or the maximum number of layers is reached;

S9, based on the projection algorithm, updates the trained output weights online according to the real-time data and calculates the real-time output.

Further, S1 specifically includes: given a set of time series data, input samples: {u(1), u(2), ..., u(n _max )} = {(y(1), u(1)), (y(2), u(2)), ..., (y(n _max ), u(n _max ))}, target output: T = {t(2), t(3), ..., t(n _max +1)}, where y and u are the system output and the controlled input respectively, and the present invention predicts the output y(n+1) at the next moment only through the current moment input (y(n), u(n)), and n _max is the number of samples. Set the maximum number of layers to S, and the maximum allowable number of neurons in each layer of the reserve pool The maximum expected output error tolerance ε, the maximum number of candidate neurons generated G _max , and the input weight distribution parameter γ = {λ _min ,λ _min +Δλ,...,λ _max }.

Furthermore, S2 specifically includes: initializing the output error vector e ₀ =T, the model output error scaling factor 0＜r＜1, assuming that each layer already has a reserve pool containing N neurons, and the network model with different numbers of layers is:

…

where j = 1, 2, ..., S, and/> is the input weight of the jth layer, the reservoir connection weight and the output weight matrix,/> is the bias matrix, x ^(j) (n) is the reservoir state vector at time n, g is the activation function, X ^(j) =[x ^(j) (1),x ^(j) (2),…x ^(j) (n)] is the reservoir state matrix, and Y ^(j) is the network output of the jth layer.

Further, S3 specifically includes: randomly selecting input weights from γ = {λ _min ,λ _min +Δλ, ...,λ _max } Reserve pool connection weight/> and bias/> Substitute the activation function (first floor) or (jth layer) Get G _max group of candidate neurons

Further, S4 specifically includes: substituting the G _max group of candidate neurons into the inequality constraints of the random configuration algorithm:

in is the output error of all reserve pool neurons in the current layers. For a single layer structure: N _sum = N, for the jth layer: /> m is the output dimension, a non-negative real number sequence/> Satisfaction/> And/> Select candidate neurons that satisfy the inequality constraints.

Further, S5 specifically includes: defining a set of variables: The candidate neurons that satisfy the inequality constraints and maximize ξ _N+1 are selected as the optimal reserve pool candidate neurons.

Further, S6 specifically includes: adding the optimal reserve pool candidate neurons to the network model, and directly calculating the output weight of the neural network model according to the target output And calculate the output root mean square error of the neural network model, use the root mean square error as the loss function, and update the output error of the model/> And the number of neurons in the reserve pool N=N+1.

Further, S7 specifically includes: if the output root mean square error of the current network model ||e ₀ || ₂ is greater than the maximum expected output error allowable value ε, and the number of neurons in the reserve pool N is less than the maximum allowable number of neurons in the reserve pool Then repeat S3 to S6; if the output root mean square error of the current network model ||e ₀ || ₂ is greater than the maximum expected output error tolerance ε, and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> Then enter the next layer to continue training, update the number of neurons in the reserve pool N = 0 and the number of reserve pool layers j = j + 1, and repeat S3 to S6; if the root mean square error ||e ₀ || ₂ of the current network model output is less than the maximum expected output error tolerance ε, or j = S and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> The training is completed and a deep network model that meets the constraints of random configuration theory is obtained.

Furthermore, S9 specifically includes: based on the projection algorithm, updating the trained output weights online according to the real-time data, and calculating the real-time output:

Where 0＜a＜1, c＞0 are two constants, g(n) is the output of the reserve pool at time n, and e(n) is the output error before correction.

Another object of the present invention is to provide a dynamic modeling system for complex systems in industrial processes using the deep recursive random configuration network industrial process accurate prediction method, comprising:

A preprocessing module, used for data preprocessing;

Initialization module, used for deep recursive random configuration network initialization;

Construct a neuron module to randomly configure weights and biases, and use a new method to construct candidate neurons in the reserve pool;

The candidate neuron selection module selects candidate neurons in a supervised manner according to inequality constraints;

The optimal neuron determination module determines the optimal candidate neuron and adds it to the reserve pool;

The output weight calculation module calculates the output weight of the current layer network by combining the target output and the reserve pool output;

The online update module is used to update the trained output weights online based on the projection algorithm and real-time data, and calculate the real-time output.

In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

First, the present invention provides a method and system for accurately predicting industrial processes using a deep recursive random configuration network. When the input order is unknown, historical information stored in a reserve pool is used for modeling. Compared with recursive neural networks and long short-term memory networks, a supervisory mechanism is introduced in the incremental construction process to randomly configure the input weights and biases of the reserve pool neurons, without the need for iterative updates through back propagation, thus avoiding the problems of being prone to local optimality, gradient disappearance or explosion caused by updating network parameters by gradient descent, and determining the network structure during the configuration process; compared with a deep echo state network, a deep recursive random configuration network can adaptively adjust weights and biases according to input data, and can theoretically guarantee the global approximation properties of the model, achieving zero error approximation of training data; compared with a single-layer recursive random configuration network, when the total number of nodes is similar, the deep version can effectively reduce and improve the training speed and prediction accuracy of the model.

Second, the present invention provides a method and system for accurately predicting industrial processes using a deep recursive random configuration network, which can form more appropriate feature representations layer by layer. Compared with a single-layer structure, it has a faster computing speed and higher prediction accuracy when the summary points are similar. Combined with the model framework designed by the present invention, it is possible to reduce the overall computational complexity of the algorithm while ensuring the prediction accuracy, and is suitable for application scenarios with high real-time requirements, and has good application prospects in the fields of industrial artificial intelligence, smart medical care, smart transportation, and unmanned driving.

For dynamic data modeling problems, most people stay at proving the existence of the global approximation ability of the model, but do not start from the perspective of constructivity. The present invention provides a dynamic modeling algorithm and system for complex systems in industrial processes, which incrementally constructs a dynamic model with global approximation ability. Moreover, the model does not require order identification, does not require gradient descent to update weights and biases, and can adaptively adjust the range of weights and biases, and is suitable for dynamic modeling problems with uncertain orders.

Third, the deep recursive random configuration network industrial process accurate prediction method has the following significant technical advances in industry:

1. Enhance model learning capabilities

Due to the deep network structure, the model can learn more complex data distribution and dynamic characteristics, thereby improving prediction accuracy, which is particularly important in complex industrial processes.

2. Dynamic and adaptive characteristics

Through loop feedback and reserve pool, as well as online weight update, the model can self-adjust and optimize during operation to adapt to dynamic changes in industrial processes.

3. Error monitoring and optimization

Through multiple layers of pooling and error feedback, the model is able to evaluate its performance in real time and adjust itself when necessary to meet accuracy requirements.

4. Save computing resources

This method is a stochastic learning algorithm in which the input weights and biases are randomly given. There is no need for gradient descent to update the weights. Compared with traditional deep learning methods, it requires fewer computing resources, which is very valuable in industrial environments with limited resources or requiring fast response.

5. Strengthen the model’s approximation capability

By selecting candidate neurons in a supervised manner through inequality constraints, the model has a universal approximation characteristic and can continuously approach the target output.

6. Hierarchical Modeling

Since the model is multi-layered, it can perform hierarchical modeling, which can not only capture more complex system dynamics characteristics, but also perform different optimizations and controls at different levels, increasing the flexibility of the model.

7. Real-time

By updating the output weights online, the model can adapt to new unknown data and conditions in real time, which is very important for industrial fields that require real-time decision-making and control.

8. Reduce maintenance costs

The adaptive and real-time updating features reduce the complexity and cost of model maintenance as the model is able to self-adjust to dynamically changing industrial environments and conditions.

The deep recursive random configuration network industrial process accurate prediction method provided by the present invention has significant industrial application potential in improving model accuracy, saving computing resources, increasing adaptive capabilities and realizing real-time control.

Fourth, the significant technical advances brought about by the industrial process accurate prediction method of the deep recursive random configuration network described in the two embodiments provided by the present invention mainly include:

1) Enhance the adaptability of the model:

By constructing a reserve pool through a supervision mechanism, determining the range of weights and biases based on input data, and using a projection algorithm to update output weights in real time, the model can dynamically adapt to system changes and improve the accuracy of modeling unknown or changing environments.

2) Improve prediction accuracy:

It adopts a data-driven weight and bias selection method instead of blind random selection, and the deep learning structure can better capture the nonlinear characteristics in the data, which significantly improves the prediction accuracy compared with traditional linear models or shallow networks.

3) Improvement of computational efficiency:

The model randomly configures weights and biases during initialization and iteration, which reduces the amount of weight calculation and improves computational efficiency.

4) Scalability and flexibility:

The hierarchical and modular design of the network structure allows the model to easily expand or modify the number of network layers and neurons according to different tasks.

5) Improvement of real-time performance:

The model can receive new data in real time and quickly adjust model parameters, making dynamic modeling more in line with actual application needs, especially in scenarios with high real-time requirements.

In the case of chemical reaction process prediction, these technological advances can achieve more stable production processes and higher product quality. In the case of power system load prediction, the scheduling and allocation of power resources can be optimized, energy waste can be reduced, and the operating efficiency of the power grid can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

FIG1 is a flow chart of a method for accurately predicting industrial processes using a deep recursive random configuration network provided by an embodiment of the present invention;

2 is a schematic diagram of the structure of a dynamic modeling system for complex systems in industrial processes provided by an embodiment of the present invention;

FIG3 is a diagram showing the effectiveness of the method proposed in the present invention on the MG system;

FIG4 is a diagram showing the fitting effects of various methods on nonlinear system identification data;

FIG5 is a diagram showing the fitting results of various methods for butane concentration data at the bottom of a debutanizer;

FIG6 is a diagram showing the effectiveness of the method of the present invention on the butane concentration data at the bottom of the debutanizer.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

Embodiment 1:

This embodiment provides a method for accurately predicting industrial processes using a deep recursive random configuration network, and its application in a complex nonlinear system dynamic modeling scenario. In this embodiment, the method is used to predict the concentration of a product in a chemical reaction process.

Data preprocessing: Collect the real-time operating data of the reactor, including parameters such as temperature, pressure, and raw material dosage, and perform normalization processing.

Initialize the network: set the initial parameters of the network, including the number of neurons in the reserve pool, learning rate, etc.

Randomly configure weights and biases: Initialize network weights and biases by random methods to construct candidate neurons in the reserve pool.

Selection of candidate neurons: Based on real-time data and system performance requirements, a genetic algorithm is used to optimize and select candidate neurons that best suit the current data characteristics.

Add the best candidate neurons: Add the selected candidate neurons to the reserve pool to form a new network structure.

Output weight calculation: Use the least squares method to calculate the output weight of the network.

Error calculation and feedback: Calculate the error based on the real-time product concentration data and network prediction value, and dynamically adjust the network structure and parameters.

By repeating the above process until the error of the network output is within an acceptable range, the dynamic modeling of the chemical reaction process is completed.

Embodiment 2:

In another embodiment, the accurate prediction method is used for power system load prediction.

The specific implementation steps are as follows:

Data preprocessing: Collect historical load data of the power system, consider the impact of weather, holidays and other factors on power load, and perform data cleaning and preprocessing.

Network initialization: Set the number of network layers and the number of neurons in each layer according to the characteristics of the power system load.

Randomly configure weights and biases: Use randomization methods to generate the initial weights and biases of the network.

Screening of candidate neurons: Use the sliding window method to dynamically select suitable candidate neurons based on past load data.

Determine the optimal candidate neurons: Use the simulated annealing algorithm to screen out candidate neurons that can maximize prediction accuracy.

Calculate output weights: Use gradient descent to optimize output weights based on target output.

Error calculation and adjustment: Perform error analysis based on the deviation between the predicted results and the actual load, and gradually optimize the network structure.

By updating and adjusting output weights in real time, the model can achieve accurate prediction of power load, thereby improving the efficiency and safety of power grid operations.

The present invention mainly improves the following problems and defects of the prior art and achieves significant technical progress:

Insufficient prediction accuracy and adaptability: Existing industrial prediction technologies may lack prediction accuracy and adaptability when processing dynamic and complex industrial data. The present invention significantly improves the prediction accuracy and adaptability by adopting a deep recursive random configuration network.

Randomness and uncertainty of weight and bias configuration: In traditional randomly configured networks, random configuration of weights and biases may lead to unstable network performance. The present invention reduces the uncertainty caused by this randomness by selecting candidate neurons in a supervised manner and optimizing the reserve pool structure.

Limitations of dynamic system modeling: Existing technologies may have limitations when modeling dynamic systems in industrial processes. The present invention utilizes a projection algorithm with real-time data updates to enhance the modeling capabilities of the model for dynamic systems.

The technical effects and significant technical progress brought about by the present invention in solving the problems of the prior art are as follows:

Through a deep recursive random configuration network and a carefully designed initialization step, the present invention significantly improves the accuracy of prediction, which is crucial for the optimization and control of complex industrial processes. Through a projection algorithm with real-time data updates, the present invention enables the network to adapt to new data and changes, providing higher adaptability. By supervised selection of optimal candidate neurons and automatic adjustment of the reserve pool and network layers, the present invention reduces the complexity of model training and adjustment. The present invention improves the depth and complexity processing capabilities of the network by repeatedly adjusting and optimizing the reserve pool neurons in a multi-layer network structure. Due to its high accuracy and adaptability, the present invention is applicable to a variety of complex industrial scenarios, improving its practicality and universality.

In view of the problems existing in the prior art, the present invention provides a method and system for accurately predicting industrial processes using a deep recursive random configuration network. The present invention is described in detail below in conjunction with the accompanying drawings.

As shown in FIG1 , the specific steps of the method proposed by the present invention are as follows:

Step 1: Data preprocessing.

Step 2: Initialization of the deep recursive random configuration network.

Set the maximum number of layers to S, and the maximum number of neurons allowed in each layer's reserve pool The initial number of neurons N=0, the maximum expected output error tolerance ε, the maximum number of candidate neurons generated G _max , the input weight distribution parameter γ={λ _min ,λ _min +Δλ,...,λ _max }, and the model output error initialization: e ₀ =T.

Step 3: Randomly select input weights from γ = {λ _min ,λ _min +Δλ,...,λ _max } Reserve pool connection weight/> and bias/> Substitute the activation function /> (first layer) or/> (jth layer) Get G _max group of candidate neurons

Step 4. Settings Substitute the G _max group of candidate neurons into the inequality constraints of the random configuration algorithm:

in is the output error of all the reserve pool neurons in the current layers (for single-layer structure: N _sum = N, for the jth layer: /> Select candidate neurons that satisfy the inequality constraints.

Step 5. Define a set of variables: Substitute the candidate neurons that satisfy the inequality constraints into the Select the candidate neuron with the largest ξ _N+1 and keep and/> Go to the next step; otherwise, λ＝λ+Δλ, when λ＜λ _max , return to step 3 to continue. If the above two conditions are not met, let τ∈(0,1-r), r＝r+τ, λ＝λ _min , and return to step 3 to continue. Add the best candidate neuron to the reserve pool to obtain the latest connection weights and biases/> and/>

Step 6: Update the neural network model, and use the least squares method to obtain the output weight of the model through the reserve pool output and the target output. Finally update the model output error/> And the number of neurons in the reserve pool N=N+1.

Step 7: When the root mean square error of the model output ||e ₀ || ₂ ≥ε, and Then return to step 3; if ||e ₀ || ₂ ≥ε, and/> Then enter the next layer to continue training, update the number of neurons in the reserve pool N = 0 and the number of reserve pool layers j = j + 1, and return to step 3. If ||e ₀ || ₂ ＜ε, or j = S and/> The deep network model training is completed.

Step 8: Based on the projection algorithm, the trained output weights are calculated according to the real-time data. Update online and calculate real-time output to analyze the generalization performance of the model.

As shown in FIG2 , the dynamic modeling system for complex systems in industrial processes provided by an embodiment of the present invention includes:

A preprocessing module, used for data preprocessing;

Initialization module, used for deep recursive random configuration network initialization;

Construct a neuron module to randomly configure weights and biases, and use a new method to construct candidate neurons in the reserve pool;

The candidate neuron selection module selects candidate neurons in a supervised manner according to inequality constraints;

The optimal neuron determination module determines the optimal candidate neuron and adds it to the reserve pool;

The output weight calculation module calculates the output weight of the current layer network by combining the target output and the reserve pool output;

The online update module is used to update the trained output weights online based on the projection algorithm and real-time data, and calculate the real-time output.

Preferably, the method for accurately predicting industrial processes using a deep recursive random configuration network provided by an embodiment of the present invention includes:

S1, data preprocessing;

S2, deep recursive random configuration network initialization;

S3, randomly configures weights and biases, and uses a new method to construct candidate neurons in the reserve pool;

S4, supervised selection of candidate neurons based on inequality constraints;

S5, determine the best candidate neurons and add them to the reserve pool;

S6, calculate the output weight of the current layer network by combining the target output and the reserve pool output;

S7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3 to S6 until the error requirement is met or the maximum allowable number of neurons in the reserve pool of the current layer is reached;

S8, when the error requirement is not met, enter the next layer and configure the reserve pool neurons of each layer in sequence according to S3 to S7 until the error requirement is met or the maximum number of layers is reached;

S9, based on the projection algorithm, updates the trained output weights online according to the real-time data and calculates the real-time output.

S1 specifically includes: given a set of time series data, input samples: {u(1),u(2),...,u(n _max )}＝{(y(1),u(1)),(y(2),u(2)),...,(y(n _max ),u(n _max ))}, target output: T＝{t(2),t(3),...,t(n _max +1)}, where y and u are the system output and the controlled input respectively. The present invention predicts the output y(n+1) at the next moment only through the current moment input (y(n),u(n)), and n _max is the number of samples. Set the maximum number of layers to S, and the maximum allowable number of neurons in each layer of the reserve pool The maximum expected output error tolerance ε, the maximum number of candidate neurons generated G _max , and the input weight distribution parameter γ = {λ _min ,λ _min +Δλ,...,λ _max }.

S2 specifically includes: initializing the output error vector e ₀ =T, the model output error scaling factor 0＜r＜1, assuming that each layer already has a reserve pool containing N neurons, and the network model with different numbers of layers is:

…

where j = 1, 2, ..., S, and/> is the input weight, reservoir connection weight and output weight of the jth layer,/> is the bias, x ^(j) (n) is the reservoir state at time n, g is the activation function, X ^(j) =[x ^(j) (1),x ^(j) (2),...x ^(j) (n)] is the reservoir state matrix, and Y ^(j) is the network output of the jth layer.

S3 specifically includes: randomly selecting an input weight matrix from γ = {λ _min ,λ _min +Δλ, ...,λ _max } Reserve pool connection weight matrix/> and bias matrix/> Substitute the activation function (first floor) or (jth layer) Get G _max group of candidate neurons

S4 specifically includes: substituting the candidate neurons of the G _max group into the inequality constraints of the random configuration algorithm:

in is the output error of all reserve pool neurons in the current layers. For a single layer structure: N _sum = N, for the jth layer: /> m is the output dimension, a non-negative real number sequence/> Satisfaction/> And/> Select candidate neurons that satisfy the inequality constraints.

S5 specifically includes: defining a set of variables: The candidate neurons that satisfy the inequality constraints and maximize ξ _N+1 are selected as the optimal reserve pool candidate neurons.

S6 specifically includes: adding the optimal reserve pool candidate neurons to the network model, and directly calculating the output weights of the neural network model according to the target output And calculate the output root mean square error of the neural network model, use the root mean square error as the loss function, and update the output error of the model/> And the number of neurons in the reserve pool N=N+1.

S7 specifically includes: if the output root mean square error of the current network model ||e ₀ || ₂ is greater than the maximum expected output error allowable value ε, and the number of neurons in the reserve pool N is less than the maximum allowable number of neurons in the reserve pool Then repeat S3 to S6; if the output root mean square error of the current network model ||e ₀ || ₂ is greater than the maximum expected output error tolerance ε, and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> Then enter the next layer to continue training, update the number of neurons in the reserve pool N = 0 and the number of reserve pool layers j = j + 1, and repeat S3 to S6; if the root mean square error ||e ₀ || ₂ of the current network model output is less than the maximum expected output error tolerance ε, or j = S and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> The training is completed and a deep network model that meets the constraints of random configuration theory is obtained.

S9 specifically includes: based on the projection algorithm, online updating of the trained output weights according to real-time data, and calculation of real-time output:

Where 0＜a＜1, c＞0 are two constants, g(n) is the output of the reserve pool at time n, and e(n) is the output error before correction.

S3 to S6 are the key points of the present invention. S3 links the input data with the weight parameters and can adaptively adjust the weight distribution to accelerate error convergence. This data-dependent parameter selection method can effectively improve the training speed and prediction accuracy of the model.

S3 adopts a new method to construct the connection weight of the reserve pool. For the N+1th newly added neuron, only the network connection weight of the newly added neuron to the original neuron is retained. and its own network connection weight The connection weight of the original neuron to the newly added neuron is set to zero/> That is, no connection:

in represents the connection weight of the ith neuron to the jth neuron,/> This construction method can ensure that when a neuron is added to the reserve pool, the output states of the first N reserve pool neurons do not change. It is only necessary to configure the new neurons that reduce the model training error to ensure the global approximation property of the network model.

The principle behind the inequality constraints of the random configuration algorithm mentioned in S4 is paraphrased as follows:

Assume that the vector space Γ on _L2 space is dense and There exists 0＜||g||＜b _g , Given 0＜r＜1 and a non-negative real number sequence/> Where/> Given the following formula:

If the random basis function The constructed output weights satisfy:

And the following inequality constraints are satisfied:

but Where T is the target output value, /> is the predicted output value of the model with N _sum +1 reservoir neurons. That is, the constructed deep neural network model has a global approximation property.

S5 is to find the candidate neurons that can reduce the model training error the fastest, thereby accelerating the convergence speed.

The calculated output weights mentioned in S6 The specific principle is reproduced as follows:

Assume that the vector space Γ on _L2 space is dense and Existence/> Given 0＜r＜1 and a non-negative real number sequence/> Where/> Given the following formula:

If the random basis function The constructed output weights satisfy:

And the following inequality constraints are satisfied:

but is the optimal output weight.

Taking the Mackey-Glass (MG) chaotic system and a nonlinear dynamic system as examples,

Example 1: The Mackey-Glass (MG) chaotic system is a very typical chaotic time series, which is generated by the following differential equation with time delay:

Among them, when τ＞16.8, the entire sequence is chaotic, has no period, and neither converges nor diverges.

Example 2: System identification is an important part of modern control theory and signal processing. The dynamic characteristics of the system are considered to be reflected in the changing input and output data. However, most actual systems are nonlinear, so nonlinear system identification becomes an important and complex problem. Given the following nonlinear system, in which u(n) = 1.05 × sin(n/45) in the training stage, the mathematical model in the test stage is as follows:

y(n+1)＝0.72y(n)+0.025y(n-1)u(n-1)+0.01u ² (n-2)+0.2u(n-3)

Example 3: The debutanizer process is an important part of the desulfurization and naphtha separation unit in the petroleum refining process. This process uses 7 auxiliary variables, namely: tower top temperature u ₁ , tower top pressure u ₂ , tower top reflux u ₃ , tower top product outflow u ₄ , sixth layer tower plate temperature u ₅ , tower bottom temperature 1u ₆ , tower bottom temperature 2u ₇ , to predict the butane concentration y at the bottom of the tower. Its nonlinear model can be described as:

y(n)＝f(u ₁ (n),u ₂ (n),u ₃ (n),u ₄ (n),u ₅ (n),u ₅ (n-1),

u ₅ (n-2),u ₅ (n-3),(u ₆ (n)+u ₇ (n))/2,

y(n-1),y(n-2),y(n-3),y(n-4))

Considering the unknown order, we reset the input of the above model, that is,

y(n)＝f(u ₁ (n),u ₂ (n),u ₃ (n),u ₄ (n),u ₅ (n),u ₆ (n),u ₇ (n),y(n-1))

In the experiment, the embodiment of the present invention selected ESN, two-layer ESN (DeepESN2), three-layer ESN (DeepESN3), RSCN and the proposed method (two-layer RSCN (DeepRSCN2), three-layer RSCN (DeepRSCN3)) for comparison. The normalized root-mean square error (NRMSE) was selected as the evaluation index of network prediction performance.

Example 1: Mackey-Glass (MG) chaotic system

Analyzing the experimental results, the training error and test error of the deep recursive random configuration network are better than those of the single-layer model and the traditional echo state network model, and the model prediction performance gradually increases with the increase in the number of layers. The following figure is a comparison of the training time of each model under different reserve pool sizes. It can be seen from Figure 3 that the deep model is still the fastest in training speed, which verifies the effectiveness of the method proposed in this invention.

Example 2: Nonlinear System Identification

Figure 4 is a diagram of the fitting effects of various methods. The fitting effect of the deep recursive random configuration network is still the best. Combined with the results in the table, we can see that DeepRSCN performs better than other models in both training and test sets, further verifying the effectiveness of the method proposed in this invention in dynamic nonlinear identification.

Example 3: Soft measurement of butane concentration at the bottom of debutanizer

Figure 5 shows the prediction results of each neural network model, from which it can be seen that DeepRSCN3 has the smallest prediction error and the model output has a higher fit with the target value. Figure 6 shows the comparison results of the training time of each model under different reserve pool sizes. The deep version has a higher training speed, which verifies the effectiveness of the method proposed in the present invention in soft measurement of industrial process parameters.

The application of the industrial process accurate prediction method of deep recursive random configuration network in industry can bring significant technological progress, including higher prediction accuracy, faster training speed, better robustness and higher resource utilization efficiency. The following are two specific application examples.

Application Example 1: Industrial Robot Fault Prediction

In the field of industrial robots, fault prediction is an important task. The deep recursive random configuration network industrial process accurate prediction method can improve the accuracy and real-time performance of fault prediction, thereby avoiding unexpected downtime and improving production efficiency.

1. Data preprocessing: Collect various sensor data of industrial robots, such as temperature, pressure, vibration, etc. Standardize or normalize the data for model training.

2. Deep recursive random configuration network initialization: Follow steps S2-S9 to perform network initialization, reserve pool neuron configuration, optimal neuron determination, network output weight calculation, and online update of real-time output.

3. Fault prediction: Based on the trained model, the real-time output of the industrial robot is calculated to predict whether a fault will occur.

This method has significant improvements in real-time performance and prediction accuracy, can detect faults earlier, avoid unexpected downtime, and improve production efficiency.

Application Example 2: Smart Grid Load Forecasting

In the field of smart grid, load forecasting is a key task. The deep recursive random configuration network industrial process accurate forecasting method can improve the accuracy and real-time performance of load forecasting, thereby achieving more effective power dispatching and improving the utilization efficiency of power resources.

1. Data preprocessing: Collect the load data of the smart grid and various factors that affect the load, such as weather conditions, holiday information, etc. Standardize or normalize the data for model training.

2. Deep recursive random configuration network initialization: Follow steps S2-S9 to perform network initialization, reserve pool neuron configuration, optimal neuron determination, network output weight calculation, and online update of real-time output.

3. Load forecasting: Based on the trained model, the real-time output of the smart grid is calculated to predict future load conditions.

This method has significant improvements in real-time performance and prediction accuracy, can achieve more effective power dispatch, and improve the utilization efficiency of power resources. It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer executable instructions and/or included in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims (4)

1. A method for accurate prediction of industrial processes based on deep recursive random configuration networks, characterized in that the method adopts an accurate prediction method based on deep recursive random configuration networks, selects and optimizes reserve pool neurons in combination with inequality constraints, and uses a projection algorithm updated by real-time data to achieve dynamic system modeling with high adaptability and high prediction accuracy; The deep recursive random configuration network industrial process accurate prediction method includes: S1, data preprocessing; S2, deep recursive random configuration network initialization; S3, randomly configures weights and biases, and uses a new method to construct candidate neurons in the reserve pool; S4, supervised selection of candidate neurons based on inequality constraints; S5, determine the best candidate neurons and add them to the reserve pool; S6, calculate the output weight of the current layer network by combining the target output and the reserve pool output; S7, calculating the error between the network output and the target output according to the output weight and the reserve pool output, and repeating S3 to S6 until the error requirement is met or the maximum allowable number of neurons in the reserve pool of the current layer is reached; S8, when the error requirement is not met, enter the next layer and configure the reserve pool neurons of each layer in sequence according to S3 to S7 until the error requirement is met or the maximum number of layers is reached; S9, based on the projection algorithm, the trained output weights are updated online according to the real-time data and the real-time output is calculated; S1 specifically includes: given a set of time series data, input samples: {(y(1),u(1)),(y(2),u(2)),...,(y(n _max ),u(n _max ))}, target output: T = {t(2),t(3),...,t(n _max +1)}, where y and u are the system output and controlled input respectively, and the output y(n+1) at the next moment is predicted by the current moment input (y(n),u(n)), and n _max is the number of samples; the maximum number of layers is set to S, and the maximum allowable number of neurons in each layer of the reserve pool Maximum expected output error tolerance ε, maximum candidate neuron generation number G _max , input weight distribution parameter γ = {λ _min ,λ _min +Δλ,...,λ _max }; S2 specifically includes: initializing the output error vector e ₀ =T, the model output error scaling factor 0＜r＜1, assuming that each layer already has a reserve pool containing N neurons, and the network model with different numbers of layers is: … where j = 1, 2, ..., S, and/> is the input weight, reservoir connection weight and output weight of the jth layer,/> is the bias of the jth layer, x ^(j) (n) is the reservoir state of the jth layer at time n, g is the activation function, X ^(j) = [x ^(j) (1), x ^(j) (2), ... x ^(j) (n)] is the reservoir state matrix, and Y ^(j) is the network output of the jth layer; S3 specifically includes: randomly selecting the weight of the input j-th layer from γ = {λ _min ,λ _min +Δλ, ...,λ _max } The weight of the reservoir connection j layer/> and the bias of the jth layer/> Substitute the activation function or Get G _max group of candidate neurons S4 specifically includes: substituting the candidate neurons of the G _max group into the inequality constraints of the random configuration algorithm: in is the output error of all reserve pool neurons in the current layers. For a single layer structure: N _sum = N, for the jth layer: m is the output dimension, a non-negative real number sequence/> Satisfaction/> and Select candidate neurons that satisfy the inequality constraints; S5 specifically includes: defining a set of variables: Select the candidate neurons that satisfy the inequality constraints and maximize ξ _N+1 as the optimal reserve pool candidate neurons; S6 specifically includes: adding the optimal reserve pool candidate neurons to the network model, and directly calculating the output weights of the neural network model according to the target output And calculate the output root mean square error of the neural network model, use the root mean square error as the loss function, and update the output error of the model/> And the number of neurons in the reserve pool N=N+1. 2. The method for accurately predicting industrial processes using a deep recursive random configuration network as described in claim 1 is characterized in that S7 specifically includes: if the output root mean square error ||e ₀ || ₂ of the current network model is greater than the maximum expected output error tolerance ε, and the number of neurons in the reserve pool N is less than the maximum allowable number of neurons in the reserve pool Then repeat S3 to S6; if the output root mean square error of the current network model ||e ₀ || ₂ is greater than the maximum expected output error tolerance ε, and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> Then enter the next layer to continue training, update the number of neurons in the reserve pool N = 0 and the number of reserve pool layers j = j + 1, and repeat S3 to S6; if the root mean square error ||e ₀ || ₂ of the current network model output is less than the maximum expected output error tolerance ε, or j = S and the number of neurons in the current layer reserve pool N is equal to the maximum allowable number of neurons in the reserve pool/> The training is completed and a deep network model that meets the constraints of random configuration theory is obtained. 3. The method for accurately predicting industrial processes using a deep recursive random configuration network according to claim 1, wherein S9 specifically comprises: based on a projection algorithm, updating the trained output weights online according to real-time data, and calculating the real-time output: Where 0＜a＜1, c＞0 are two constants, g(n) is the output of the reserve pool at time n, and e(n) is the output error before correction. 4. A dynamic modeling system using the industrial process accurate prediction method of the deep recursive random configuration network as claimed in any one of claims 1 to 3, characterized in that it comprises: A preprocessing module, used for data preprocessing; Initialization module, used for deep recursive random configuration network initialization; Construct a neuron module to randomly configure weights and biases, and use a new method to construct candidate neurons in the reserve pool; The candidate neuron selection module selects candidate neurons in a supervised manner according to inequality constraints; The optimal neuron determination module determines the optimal candidate neuron and adds it to the reserve pool; The output weight calculation module calculates the output weight of the current layer network by combining the target output and the reserve pool output; The online update module is used to update the trained output weights online based on the projection algorithm and real-time data, and calculate the real-time output.